bwa-mem3 is a short-read aligner derived from bwa-mem2, carrying correctness fixes, performance improvements, and new features (methylation alignment, shared-memory index, mimalloc allocator) maintained by Fulcrum Genomics.
Wall-clock speedup of the current release (v0.18.0*) against bwa 0.7.19, bwa-mem2 v2.2.1, and minibwa, on the wgs-5M sample. Cells are stock / --fast.
| arch | wall_s | vs bwa | vs bwa-mem2 | vs minibwa |
|---|---|---|---|---|
| ARM | 28.42 / 12.44 | 8.82x / 20.15x | — | 1.43x / 3.27x |
| x86 | 15.44 / 8.95 | 12.85x / 22.16x | 5.72x / 9.87x | 1.94x / 3.36x |
* v0.18.0, like v0.13.0 through v0.17.0, is measured with a denser suffix-array index — stride 2 instead of the stock stride 8 — built with bwa-mem3 re-sa -u 1 (or staged with bwa-mem3 shm -u 1), both new in v0.13.0. Alignment records are byte-identical by construction (each added suffix-array sample is computed from the on-disk ones), and measured so; only the @PG header line differs, since its command line names the index path: a full-content fgumi compare bams of wgs-5M (10,030,558 records) aligned on the stride-2 index against the stock stride-8 index — v0.13.0 against the output-identical v0.12.0 — found 0 differences on both benchmark hosts, Graviton4 m8g (NEON) and AMD m8a (AVX-512), and the change that added it (#510) found 0 differences in a record-by-record comparison of core SAM fields and tags on 5M-pair hg38 WGS and WES slices (that PR does not record the host or SIMD tier). It costs ~12 GB more memory for the human genome. v0.13.0 through v0.18.0 are all measured on it in the tables below, so v0.18.0's vs prev is a like-for-like comparison. Every older release and every comparator uses the stock index.
Tip
📈 Full release-history table — every bwa-mem3 release since v0.2.1, full methodology, and version pins.
Click to expand
Graviton4 (m8g, arm64/NEON)
| release | wall_s | vs bwa | vs bwa-mem2 | vs minibwa | vs prev |
|---|---|---|---|---|---|
| bwa | 250.66 | 1.00x | — | 0.16x | — |
| bwa-mem2 | — | — | — | — | — |
| minibwa | 40.66 | 6.16x | — | 1.00x | — |
| v0.2.1 | 120.63 | 2.08x | — | 0.34x | — |
| v0.2.2 | 118.26 | 2.12x | — | 0.34x | 1.020x |
| v0.3.0 | 102.35 | 2.45x | — | 0.40x | 1.155x |
| v0.4.0 | 104.10 | 2.41x | — | 0.39x | 0.983x |
| v0.5.0 | 104.20 / 36.73 | 2.41x / 6.82x | — | 0.39x / 1.11x | 0.999x |
| v0.6.0 | 97.40 / 36.22 | 2.57x / 6.92x | — | 0.42x / 1.12x | 1.070x / 1.014x |
| v0.7.0 | 91.45 / 39.33 | 2.74x / 6.37x | — | 0.44x / 1.03x | 1.065x / 0.921x |
| v0.8.0 | 76.42 / 27.71 | 3.28x / 9.05x | — | 0.53x / 1.47x | 1.197x / 1.419x |
| v0.9.0 | 76.39 / 27.85 | 3.28x / 9.00x | — | 0.53x / 1.46x | 1.000x / 0.995x |
| v0.10.0 | 70.85 / 27.04 | 3.54x / 9.27x | — | 0.57x / 1.50x | 1.078x / 1.030x |
| v0.11.0 | 63.96 / 25.52 | 3.92x / 9.82x | — | 0.64x / 1.59x | 1.108x / 1.060x |
| v0.12.0 | 58.05 / 18.70 | 4.32x / 13.40x | — | 0.70x / 2.17x | 1.102x / 1.365x |
| v0.13.0* | 54.32 / 18.10 | 4.61x / 13.85x | — | 0.75x / 2.25x | 1.069x / 1.033x |
| v0.14.0* | 41.50 / 17.17 | 6.04x / 14.60x | — | 0.98x / 2.37x | 1.309x / 1.054x |
| v0.14.1* | 35.08 / 15.93 | 7.15x / 15.74x | — | 1.16x / 2.55x | 1.183x / 1.078x |
| v0.15.0* | 34.16 / 15.42 | 7.34x / 16.26x | — | 1.19x / 2.64x | 1.027x / 1.033x |
| v0.15.1* | 31.40 / 14.12 | 7.98x / 17.75x | — | 1.29x / 2.88x | 1.088x / 1.092x |
| v0.16.0* | 28.81 / 12.64 | 8.70x / 19.83x | — | 1.41x / 3.22x | 1.090x / 1.117x |
| v0.17.0* | 28.39 / 12.48 | 8.83x / 20.08x | — | 1.43x / 3.26x | 1.015x / 1.013x |
| v0.18.0* | 28.42 / 12.44 | 8.82x / 20.15x | — | 1.43x / 3.27x | 0.999x / 1.003x |
AMD (m8a, x86-64/AVX-512)
| release | wall_s | vs bwa | vs bwa-mem2 | vs minibwa | vs prev |
|---|---|---|---|---|---|
| bwa | 198.33 | 1.00x | 0.45x | 0.15x | — |
| bwa-mem2 | 88.30 | 2.25x | 1.00x | 0.34x | — |
| minibwa | 30.03 | 6.60x | 2.94x | 1.00x | — |
| v0.2.1 | 54.20 | 3.66x | 1.63x | 0.55x | 1.629x |
| v0.2.2 | 52.18 | 3.80x | 1.69x | 0.58x | 1.039x |
| v0.3.0 | 49.12 | 4.04x | 1.80x | 0.61x | 1.062x |
| v0.4.0 | 51.04 | 3.89x | 1.73x | 0.59x | 0.962x |
| v0.5.0 | 50.92 / 22.08 | 3.89x / 8.98x | 1.73x / 4.00x | 0.59x / 1.36x | 1.002x |
| v0.6.0 | 49.56 / 21.63 | 4.00x / 9.17x | 1.78x / 4.08x | 0.61x / 1.39x | 1.027x / 1.021x |
| v0.7.0 | 47.71 / 21.89 | 4.16x / 9.06x | 1.85x / 4.03x | 0.63x / 1.37x | 1.039x / 0.988x |
| v0.8.0 | 42.78 / 17.76 | 4.64x / 11.17x | 2.06x / 4.97x | 0.70x / 1.69x | 1.115x / 1.233x |
| v0.9.0 | 42.89 / 17.67 | 4.62x / 11.22x | 2.06x / 5.00x | 0.70x / 1.70x | 0.997x / 1.005x |
| v0.10.0 | 42.30 / 17.63 | 4.69x / 11.25x | 2.09x / 5.01x | 0.71x / 1.70x | 1.014x / 1.002x |
| v0.11.0 | 41.07 / 16.80 | 4.83x / 11.81x | 2.15x / 5.26x | 0.73x / 1.79x | 1.030x / 1.049x |
| v0.12.0 | 32.32 / 13.92 | 6.14x / 14.25x | 2.73x / 6.34x | 0.93x / 2.16x | 1.271x / 1.207x |
| v0.13.0* | 29.24 / 13.24 | 6.78x / 14.98x | 3.02x / 6.67x | 1.03x / 2.27x | 1.105x / 1.051x |
| v0.14.0* | 23.17 / 12.12 | 8.56x / 16.36x | 3.81x / 7.29x | 1.30x / 2.48x | 1.262x / 1.092x |
| v0.14.1* | 21.39 / 11.93 | 9.27x / 16.62x | 4.13x / 7.40x | 1.40x / 2.52x | 1.083x / 1.016x |
| v0.15.0* | 20.30 / 11.11 | 9.77x / 17.85x | 4.35x / 7.95x | 1.48x / 2.70x | 1.054x / 1.074x |
| v0.15.1* | 18.44 / 10.38 | 10.76x / 19.11x | 4.79x / 8.51x | 1.63x / 2.89x | 1.101x / 1.070x |
| v0.16.0* | 15.96 / 9.06 | 12.43x / 21.89x | 5.53x / 9.75x | 1.88x / 3.31x | 1.155x / 1.146x |
| v0.17.0* | 15.25 / 9.04 | 13.01x / 21.94x | 5.79x / 9.77x | 1.97x / 3.32x | 1.047x / 1.002x |
| v0.18.0* | 15.44 / 8.95 | 12.85x / 22.16x | 5.72x / 9.87x | 1.94x / 3.36x | 0.988x / 1.010x |
vs prev is the release-over-release speedup (prev_wall / this_wall, >1 = faster) vs the previous release on this same host, stock / --fast. Since v0.16.0 every bwa-mem3 release in these tables, old ones included, writes BAM natively (--bam=0) while bwa, bwa-mem2 and minibwa pipe SAM through samtools view -u; earlier tables sent every arm through SAM, so vs prev stays like-for-like but the comparator ratios are not comparable across that change. Since v0.17.0 every bwa-mem3 release whose Makefile supports link-time optimization (LTO), v0.15.1 onward, is built with it, old ones included; releases before v0.15.1 predate it and are built without. Every row was re-measured, so neither absolute times nor ratios are comparable with earlier versions of this table. The first release's predecessor is upstream bwa-mem2 — bwa-mem3 is its successor — so v0.2.1's vs prev is its speedup over bwa-mem2 (blank on ARM, where upstream has no build).
Version pins: bwa 0.7.19 · bwa-mem2 v2.2.1 · minibwa commit d6d9f87d (minibwa-0.7). "ARM" = Graviton4 m8g (arm64/NEON, no SMT); "x86" = AMD m8a (x86-64/AVX-512, no SMT). Both are the general-purpose siblings of the c8g/c8a hosts used through v0.12.0 — same CPU family, same core count, no SMT, but 4 GiB/vCPU instead of 2 so every historical arm fits in memory — so absolute times are not comparable with earlier versions of this table, only ratios within it. (The x86 arm replaced an earlier Intel c7i arm, which ran 16 vCPUs over 8 physical cores under 2-way SMT and so wasn't a real core-for-core match for Graviton's 16 real cores.) No ARM bwa-mem2 build exists, hence the blank cells there. Every arm for a given arch ran interleaved on one fixed on-demand host — 5 rounds for v0.13.0 through v0.18.0, 2 for the rest, median wall-clock shown — so these are same-host comparisons, not medians pooled across separate runs. — means the release predates the comparator or predates --fast. * v0.13.0 through v0.18.0 rows use the stride-2 suffix-array index described above; all other rows use the stock stride-8 index. Regenerated at each release; see Benchmarks.
Warning
--fast is not alignment-identical to the default preset — it trades some sensitivity/specificity at the extremes (repetitive/multi-mapping regions, low-MAPQ reads) for the speedup above. See "Three ways to run it" below before switching a production pipeline to it.
bwa-mem3 has three alignment modes that differ in what alignments come out, not just in speed:
| mode | where reads align | when to use |
|---|---|---|
| plain (default) | bwa-mem2's alignments plus bonafide correctness fixes, with two extra tags (MQ:i, HN:i) and an enriched header. On the cells re-measured for release 0.7.1, the complete alignment-record stream (tags stripped) is byte-identical to bwa-mem2 v2.2.1 on wgs-5M/wes-5M/hic-1M (x86 c6a AVX2, with a c6a/c8g cross-arch check confirming the Arm c8g NEON build matches) — differing only by those additive tags and the header. Separately, a 1.07M-record HG00096 WGS slice shows zero diverging primary alignments (x86, primary-only; not part of the cross-arch or complete-stream checks). |
Migrating a pipeline, validating against bwa/bwa-mem2, or any new pipeline. |
--compat=bwa-mem2 / --compat=bwa-mem |
Byte-for-byte identical alignment records to a specific upstream (bwa-mem2 v2.2.1 or bwa 0.7.19), @PG excluded and -t/-K matched. The two targets are not interchangeable. |
Diff-clean validation against an existing bwa/bwa-mem2 golden. |
--fast |
Faster, and not record-compatible with the default: it reshuffles the low-confidence tail (~85% of the reads it re-places had MAPQ 0; the confident MAPQ-60 core moves on ≤0.5%, 0.011% on wgs-5M) while staying accuracy-neutral against golden truth (≤0.02 pp across the WGS and methylation sims). Figures from the benchmark release-validation cells (wgs-5M/wes-5M/panel-twist-5M at 5 M reads, hic-1M/sbx-1M at 1 M) across every SIMD tier (AVX2 c6a, AVX-512 c7a/c7i, NEON c7g/c8g; meth on m7i), each a .4xlarge host at -t 16, -K 160000000. |
High-throughput pipelines where you care about the confident, uniquely-mapped calls. |
--compat is mutually exclusive with --fast (and with --meth and --proper-pair-from-emitted). See Alignment modes for the full side-by-side and Equivalence with bwa-mem2 for the field-by-field audit.
By default bwa-mem3 keeps bwa-mem2's command-line defaults, so it drops into an existing pipeline unchanged. For the fastest configuration — and what each recommended deviation from the bwa defaults trades for speed — see Settings profiles: bwa drop-in vs recommended.
Full documentation: https://bwa-mem3.readthedocs.io
The recommended way to install bwa-mem3 is via bioconda:
mamba install -c bioconda bwa-mem3
bwa-mem3 versionPrebuilt packages are available for linux-64, linux-aarch64, and osx-arm64.
git clone --recursive https://github.com/fg-labs/bwa-mem3.git
cd bwa-mem3
make
./bwa-mem3 versionSee the installation guide for prerequisites and architecture-specific notes.
- Benchmarks — published results for every release, and how they are measured
- bwa-mem3-rs — Rust bindings for bwa-mem3
- bioconda recipe — conda package on bioconda
- fgumi — UMI-aware consensus and deduplication
- bwa-mem2 — upstream project
MIT. See the License page in the docs.
Please cite the bwa-mem2 paper (Vasimuddin Md et al., IPDPS 2019). See the Citation page for BibTeX.
File issues and pull requests on fg-labs/bwa-mem3.